Add virtualized XSAVE support to bhyve which permits guests to use XSAVE and

XSAVE-enabled features like AVX.
- Store a per-cpu guest xcr0 register.  When switching to the guest FPU
  state, switch to the guest xcr0 value.  Note that the guest FPU state is
  saved and restored using the host's xcr0 value and xcr0 is saved/restored
  "inside" of saving/restoring the guest FPU state.
- Handle VM exits for the xsetbv instruction by updating the guest xcr0.
- Expose the XSAVE feature to the guest only if the host has enabled XSAVE,
  and only advertise XSAVE features enabled by the host to the guest.
  This ensures that the guest will only adjust FPU state that is a subset
  of the guest FPU state saved and restored by the host.

Reviewed by:	grehan

sys/amd64/vmm/intel/vmx.c

@@ -1227,6 +1227,40 @@ vmx_clear_nmi_blocking(struct vmx *vmx, int vcpuid)
 	vmcs_write(VMCS_GUEST_INTERRUPTIBILITY, gi);
 }
 
+static int
+vmx_emulate_xsetbv(struct vmx *vmx, int vcpu)
+{
+	struct vmxctx *vmxctx;
+	uint64_t xcrval;
+	const struct xsave_limits *limits;
+
+	vmxctx = &vmx->ctx[vcpu];
+	limits = vmm_get_xsave_limits();
+
+	/* We only handle xcr0 if the host has XSAVE enabled. */
+	if (vmxctx->guest_rcx != 0 || !limits->xsave_enabled)
+		return (UNHANDLED);
+
+	xcrval = vmxctx->guest_rdx << 32 | (vmxctx->guest_rax & 0xffffffff);
+	if ((xcrval & ~limits->xcr0_allowed) != 0)
+		return (UNHANDLED);
+
+	if (!(xcrval & XFEATURE_ENABLED_X87))
+		return (UNHANDLED);
+
+	if ((xcrval & (XFEATURE_ENABLED_AVX | XFEATURE_ENABLED_SSE)) ==
+	    XFEATURE_ENABLED_AVX)
+		return (UNHANDLED);
+
+	/*
+	 * This runs "inside" vmrun() with the guest's FPU state, so
+	 * modifying xcr0 directly modifies the guest's xcr0, not the
+	 * host's.
+	 */
+	load_xcr(0, xcrval);
+	return (HANDLED);
+}
+
 static int
 vmx_emulate_cr_access(struct vmx *vmx, int vcpu, uint64_t exitqual)
 {
@@ -1774,6 +1808,9 @@ vmx_exit_process(struct vmx *vmx, int vcpu, struct vm_exit *vmexit)
 		vlapic = vm_lapic(vmx->vm, vcpu);
 		handled = vmx_handle_apic_write(vlapic, qual);
 		break;
+	case EXIT_REASON_XSETBV:
+		handled = vmx_emulate_xsetbv(vmx, vcpu);
+		break;
 	default:
 		vmm_stat_incr(vmx->vm, vcpu, VMEXIT_UNKNOWN, 1);
 		break;

sys/amd64/vmm/vmm.c

@@ -89,6 +89,7 @@ struct vcpu {
 	struct vlapic	*vlapic;
 	int		 vcpuid;
 	struct savefpu	*guestfpu;	/* guest fpu state */
+	uint64_t	guest_xcr0;
 	void		*stats;
 	struct vm_exit	exitinfo;
 	enum x2apic_state x2apic_state;
@@ -206,6 +207,7 @@ vcpu_init(struct vm *vm, uint32_t vcpu_id)
 	vcpu->vcpuid = vcpu_id;
 	vcpu->vlapic = VLAPIC_INIT(vm->cookie, vcpu_id);
 	vm_set_x2apic_state(vm, vcpu_id, X2APIC_ENABLED);
+	vcpu->guest_xcr0 = XFEATURE_ENABLED_X87;
 	vcpu->guestfpu = fpu_save_area_alloc();
 	fpu_save_area_reset(vcpu->guestfpu);
 	vcpu->stats = vmm_stat_alloc();
@@ -815,6 +817,10 @@ restore_guest_fpustate(struct vcpu *vcpu)
 	fpu_stop_emulating();
 	fpurestore(vcpu->guestfpu);
 
+	/* restore guest XCR0 if XSAVE is enabled in the host */
+	if (rcr4() & CR4_XSAVE)
+		load_xcr(0, vcpu->guest_xcr0);
+
 	/*
 	 * The FPU is now "dirty" with the guest's state so turn on emulation
 	 * to trap any access to the FPU by the host.
@@ -829,6 +835,12 @@ save_guest_fpustate(struct vcpu *vcpu)
 	if ((rcr0() & CR0_TS) == 0)
 		panic("fpu emulation not enabled in host!");
 
+	/* save guest XCR0 and restore host XCR0 */
+	if (rcr4() & CR4_XSAVE) {
+		vcpu->guest_xcr0 = rxcr(0);
+		load_xcr(0, vmm_get_host_xcr0());
+	}
+
 	/* save guest FPU state */
 	fpu_stop_emulating();
 	fpusave(vcpu->guestfpu);

sys/amd64/vmm/vmm_host.c

@@ -38,11 +38,14 @@ __FBSDID("$FreeBSD$");
 
 #include "vmm_host.h"
 
-static uint64_t vmm_host_efer, vmm_host_pat, vmm_host_cr0, vmm_host_cr4;
+static uint64_t vmm_host_efer, vmm_host_pat, vmm_host_cr0, vmm_host_cr4,
+	vmm_host_xcr0;
+static struct xsave_limits vmm_xsave_limits;
 
 void
 vmm_host_state_init(void)
 {
+	int regs[4];
 
 	vmm_host_efer = rdmsr(MSR_EFER);
 	vmm_host_pat = rdmsr(MSR_PAT);
@@ -57,6 +60,21 @@ vmm_host_state_init(void)
 	vmm_host_cr0 = rcr0() | CR0_TS;
 
 	vmm_host_cr4 = rcr4();
+
+	/*
+	 * Only permit a guest to use XSAVE if the host is using
+	 * XSAVE.  Only permit a guest to use XSAVE features supported
+	 * by the host.  This ensures that the FPU state used by the
+	 * guest is always a subset of the saved guest FPU state.
+	 */
+	if (vmm_host_cr4 & CR4_XSAVE) {
+		vmm_xsave_limits.xsave_enabled = 1;
+		vmm_host_xcr0 = rxcr(0);
+		vmm_xsave_limits.xcr0_allowed = vmm_host_xcr0;
+
+		cpuid_count(0xd, 0x0, regs);
+		vmm_xsave_limits.xsave_max_size = regs[1];
+	}
 }
 
 uint64_t
@@ -87,6 +105,13 @@ vmm_get_host_cr4(void)
 	return (vmm_host_cr4);
 }
 
+uint64_t
+vmm_get_host_xcr0(void)
+{
+
+	return (vmm_host_xcr0);
+}
+
 uint64_t
 vmm_get_host_datasel(void)
 {
@@ -122,3 +147,10 @@ vmm_get_host_idtrbase(void)
 
 	return (r_idt.rd_base);
 }
+
+const struct xsave_limits *
+vmm_get_xsave_limits(void)
+{
+
+	return (&vmm_xsave_limits);
+}

sys/amd64/vmm/vmm_host.h

@@ -33,17 +33,25 @@
 #error "no user-servicable parts inside"
 #endif
 
+struct xsave_limits {
+	int		xsave_enabled;
+	uint64_t	xcr0_allowed;
+	uint32_t	xsave_max_size;
+};
+
 void vmm_host_state_init(void);
 
 uint64_t vmm_get_host_pat(void);
 uint64_t vmm_get_host_efer(void);
 uint64_t vmm_get_host_cr0(void);
 uint64_t vmm_get_host_cr4(void);
+uint64_t vmm_get_host_xcr0(void);
 uint64_t vmm_get_host_datasel(void);
 uint64_t vmm_get_host_codesel(void);
 uint64_t vmm_get_host_tsssel(void);
 uint64_t vmm_get_host_fsbase(void);
 uint64_t vmm_get_host_idtrbase(void);
+const struct xsave_limits *vmm_get_xsave_limits(void);
 
 /*
  * Inline access to host state that is used on every VM entry

sys/amd64/vmm/x86.c

@@ -30,17 +30,19 @@
 __FBSDID("$FreeBSD$");
 
 #include <sys/param.h>
-#include <sys/types.h>
+#include <sys/pcpu.h>
 #include <sys/systm.h>
 #include <sys/cpuset.h>
 
 #include <machine/clock.h>
 #include <machine/cpufunc.h>
 #include <machine/md_var.h>
+#include <machine/segments.h>
 #include <machine/specialreg.h>
 
 #include <machine/vmm.h>
 
+#include "vmm_host.h"
 #include "x86.h"
 
 #define	CPUID_VM_HIGH		0x40000000
@@ -53,6 +55,8 @@ int
 x86_emulate_cpuid(struct vm *vm, int vcpu_id,
 		  uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
 {
+	const struct xsave_limits *limits;
+	uint64_t cr4;
 	int error, enable_invpcid;
 	unsigned int 	func, regs[4];
 	enum x2apic_state x2apic_state;
@@ -147,11 +151,27 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id,
 				regs[2] |= CPUID2_X2APIC;
 
 			/*
-			 * Hide xsave/osxsave/avx until the FPU save/restore
-			 * issues are resolved
+			 * Only advertise CPUID2_XSAVE in the guest if
+			 * the host is using XSAVE.
 			 */
-			regs[2] &= ~(CPUID2_XSAVE | CPUID2_OSXSAVE |
-				     CPUID2_AVX);
+			if (!(regs[2] & CPUID2_OSXSAVE))
+				regs[2] &= ~CPUID2_XSAVE;
+
+			/*
+			 * If CPUID2_XSAVE is being advertised and the
+			 * guest has set CR4_XSAVE, set
+			 * CPUID2_OSXSAVE.
+			 */
+			regs[2] &= ~CPUID2_OSXSAVE;
+			if (regs[2] & CPUID2_XSAVE) {
+				error = vm_get_register(vm, vcpu_id,
+				    VM_REG_GUEST_CR4, &cr4);
+				if (error)
+					panic("x86_emulate_cpuid: error %d "
+					      "fetching %%cr4", error);
+				if (cr4 & CR4_XSAVE)
+					regs[2] |= CPUID2_OSXSAVE;
+			}
 
 			/*
 			 * Hide monitor/mwait until we know how to deal with
@@ -219,7 +239,6 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id,
 
 		case CPUID_0000_0006:
 		case CPUID_0000_000A:
-		case CPUID_0000_000D:
 			/*
 			 * Handle the access, but report 0 for
 			 * all options
@@ -240,6 +259,57 @@ x86_emulate_cpuid(struct vm *vm, int vcpu_id,
 			regs[3] = vcpu_id;
 			break;
 
+		case CPUID_0000_000D:
+			limits = vmm_get_xsave_limits();
+			if (!limits->xsave_enabled) {
+				regs[0] = 0;
+				regs[1] = 0;
+				regs[2] = 0;
+				regs[3] = 0;
+				break;
+			}
+
+			cpuid_count(*eax, *ecx, regs);
+			switch (*ecx) {
+			case 0:
+				/*
+				 * Only permit the guest to use bits
+				 * that are active in the host in
+				 * %xcr0.  Also, claim that the
+				 * maximum save area size is
+				 * equivalent to the host's current
+				 * save area size.  Since this runs
+				 * "inside" of vmrun(), it runs with
+				 * the guest's xcr0, so the current
+				 * save area size is correct as-is.
+				 */
+				regs[0] &= limits->xcr0_allowed;
+				regs[2] = limits->xsave_max_size;
+				regs[3] &= (limits->xcr0_allowed >> 32);
+				break;
+			case 1:
+				/* Only permit XSAVEOPT. */
+				regs[0] &= CPUID_EXTSTATE_XSAVEOPT;
+				regs[1] = 0;
+				regs[2] = 0;
+				regs[3] = 0;
+				break;
+			default:
+				/*
+				 * If the leaf is for a permitted feature,
+				 * pass through as-is, otherwise return
+				 * all zeroes.
+				 */
+				if (!(limits->xcr0_allowed & (1ul << *ecx))) {
+					regs[0] = 0;
+					regs[1] = 0;
+					regs[2] = 0;
+					regs[3] = 0;
+				}
+				break;
+			}
+			break;
+
 		case 0x40000000:
 			regs[0] = CPUID_VM_HIGH;
 			bcopy(bhyve_id, &regs[1], 4);